Dynamo: Amazon s Highly Available Key-Value Store

Transcription

1 Dynamo: Amazon s Highly Available Key-Value Store DeCandia et al. Amazon.com Presented by Sushil CS

2 Motivation A storage system that attains high availability, performance and durability Decentralized techniques Data partitioning Replica synchronization Membership 2

3 Agenda Introduction System Architecture Implementation Experiments Conclusion 3

4 Amazon.com One of the largest e-business platform 3 million checkouts on a peak day Also a major cloud hosting service Customer trust Hundreds of thousands of machines Network failures/ disk failures is a norm Availablility? 4

5 Service Oriented Architecture Loosely coupled replicated services Stateless Persistent store Services e.g - Recommendation, top selling, catalog, etc 5

6 Service Requirements Query Model key, value Code versioning systems Must be able to make tradeoffs between availability, consistency, durability percentile SLA Example Shopping Cart service why not relational database? 6

7 Design Considerations Highly available (writes) Eventually consistent Merge during read Handled by applications Less manual interaction Incrementally scalable Completely decentralized Contrast Bigtable? 7

8 Challenges Partitioning Availability (writes) Handling Failures Temporary Permanent Membership 8

9 Partitioning Consistent Hashing Contrast linear hashing? MD5 hash Replication N nodes Preference list Multiple data centers Coordinator Is this a global view? Hierarchical namespace? 9

10 Ring partitioning Problems Non uniform data Heterogeneity Use virtual nodes Multiple tokens per node Add/remove node keeps system load steady Incorrect buckets are bounded. 10

11 Data Versioning Asynchronous updates Merging maintain new immutable version Objects as blobs Syntactic and semantic Associative and commutative? Multiple branches Reconcile versions during read Last write wins Vector clocks 11

12 Vector clocks list (node, counter) Partial ordering Merged during read 12

13 Get and Put API get(key) returns list<object> put(key, context, object) Context contains metadata & version Load balancer or client library Request forwarding to coordinator Quorum durability and anti-entropy R nodes for read W nodes for write 13

14 Hinted Handoff Sloppy quorum Use first N healthy nodes N=3, B unresponsive Sent to E, metadata hints B 14

15 Replica Synchronization Anti-entropy Merkle trees Leaf hash value of key Parents - hash of childs One tree per key range/virtual node Peers compare merkle trees Advantages Less reads 15

16 Gossip Admin issue command to join/remove node Serving node records in its local membership history Gossip based protocol used to agree on the memberships Partition and Placement information sent during gossip 16

17 Failure detection 17

18 Implementation Local persistence BerkleyDataBase Transactional Data Store Request Coordination SEDA architecture 18

19 Read Operation Send read requests to nodes Wait for minimum no of responses (R) Too few replies fail within time bound Gather and find conflicting versions Create context (opaque to caller) Read repair 19

20 Values of N, R and W N represents durability Typical value 3 W and R affect durability, availability, consistency What if W is low? Durability and Availability go hand-inhand? 20

21 Results 21

22 Out of balance nodes 22

23 Partition Strategies T random tokens per node and partition by token value Scan a range Updating merkle trees T random tokens per node and equal partitions Decoupling partition and placement Changing the placement scheme at runtime Q/S tokens per node, equal partitions 23

24 Conclusion Dynamo has provided high availability and fault tolerance Provides owners to customize according to their SLA requirements Decentralized techniques can provide highly available system 24

25 Current State Some of the principles used by S3 Open source implementation Cassandra Voldemort 25